Abstract
Story retelling is a fundamental medium for the transmission of information between individuals and among social groups. Besides conveying factual information, stories also contain affective information. Though natural language processing techniques have advanced considerably in recent years, the extent to which machines can be trained to identify and track emotions across retellings is unknown. This study leverages the powerful RoBERTa model, based on a transformer architecture, to derive emotion-rich story embeddings from a unique dataset of 25,728 story retellings. The initial stories were centered around five emotional events (joy, sadness, embarrassment, risk, and disgust-though the stories did not contain these emotion words) and three intensities (high, medium, and low). Our results indicate (1) that RoBERTa can identify emotions in stories it was not trained on, (2) that the five emotions and their intensities are preserved when they are transmitted in the form of retellings, (3) that the emotions in stories are increasingly well-preserved as they experience additional retellings, and (4) that among the five emotions, risk and disgust are least well-preserved, compared with joy, sadness, and embarrassment. This work is a first step toward quantifying situation-driven emotions with machines.